Description

This study aims to externally validate the CLOVER score, a machine learning-based predictive model designed to identify patients with venous thromboembolism (VTE) who are at increased risk of having an occult cancer. The study combines a retrospective and a prospective cohort.

Retrospective cohort:

Patients aged ≥18 years with objectively confirmed symptomatic deep vein thrombosis (DVT) and/or pulmonary embolism (PE) diagnosed between January 1, 2000, and August 31, 2022, in participating hospitals across Spain will be included. Patients from centers involved in the model derivation (Hospital Universitario Infanta Leonor and Hospital Universitario de Fuenlabrada) will be excluded. Cases are defined as patients with a histologically confirmed cancer diagnosis occurring between 1 and 24 months after the index VTE event. Controls are defined as patients without a cancer diagnosis during the same period. All variables required for the CLOVER model (age, D-dimer, systolic blood pressure, ALT, hemoglobin, creatinine, total cholesterol, platelet count, triglycerides, leukocyte count, weight, chronic lung disease, heart rate, sex, and previous VTE recurrence) will be extracted from electronic health records using a standardized data collection form. The CLOVER model will be applied to each patient to assess its discrimination for cancer prediction in the retrospective cohort.

Prospective cohort:

From December 1, 2025, to December 31, 2027, consecutive adult patients with objectively confirmed symptomatic VTE will be recruited in participating hospitals. All patients will undergo a standard clinical evaluation including medical history, physical examination, basic laboratory testing (complete blood count and biochemistry), chest X-ray, and age- and sex-appropriate cancer screening tests according to clinical practice guidelines. The CLOVER score will be calculated at the time of VTE diagnosis using a dedicated web-based tool (CLOVER-Web). Patients will be classified as "low risk" or "high risk" based on the optimal F1-score threshold (0.487), corresponding to a sensitivity of 51%, specificity of 95%, PPV of 46%, and NPV of 96%. All participants will be followed for at least two years to determine whether an occult cancer is diagnosed. Further diagnostic testing for suspected cancer will be performed at the discretion of the treating physician, regardless of CLOVER result.

Model performance and bias assessment:

The external validation will evaluate model discrimination using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and the area under the ROC curve (AUC). Additional analyses will assess potential prediction bias across demographic and clinical subgroups (e.g., age, sex, comorbidities). Model calibration and temporal performance will also be reviewed, with predefined procedures for identifying deviations from expected behavior. Weekly reports during the prospective phase will alert investigators to performance drift or potential implementation issues in the web-based tool.

Clinician satisfaction:

A secondary objective is to evaluate clinician satisfaction with the CLOVER-Web tool. An online questionnaire specifically designed for this study will be administered to participating clinicians to assess usability, clarity, and clinical utility.

Ethical considerations:

The study follows the Declaration of Helsinki and Spanish regulations on biomedical research and data protection. Retrospective data collection will request exemption from informed consent due to the use of anonymized clinical information. Prospective participants will provide written or electronic informed consent. All data will be pseudonymized and stored securely.